Inference apparatus, training apparatus, and inference method

Abstract

An inference apparatus includes one or more memories; and one or more processors. The one or more processors configured to acquire a latent state from input data regarding a control target; acquire a future latent state from the latent state and control data; infer, from the future latent state, a time series of a task to be executed by the control target to be controlled based on the control data; calculate a loss between the time series of the task and data indicating a target state; and update the control data based on the loss.

1 . An inference apparatus, comprising:
 one or more memories; and   one or more processors configured to:
 acquire a latent state from input data regarding a control target; 
 acquire a future latent state from the latent state and control data; 
 infer, from the future latent state, a time series of a task to be executed by the control target to be controlled based on the control data; 
 calculate a loss between the time series of the task and data indicating a target state; and 
 update the control data based on the loss. 
   
     
     
         2 . The inference apparatus according to  claim 1 , wherein
 the one or more processors acquire the latent state, acquire the future latent state, and infer the time series of the task through a trained neural network model.   
     
     
         3 . The inference apparatus according to  claim 2 , wherein
 the neural network model includes a fully connected layer, and   the one or more processors acquire the latent state based on an output result of the fully connected layer.   
     
     
         4 . The inference apparatus according to  claim 2 , wherein
 the one or more processors acquire the latent state by using information regarding a joint of the control target as the input data.   
     
     
         5 . The inference apparatus according to  claim 2 , wherein
 the one or more processors acquire the latent state by using information regarding an image of the control target or an optical flow as the input data.   
     
     
         6 . The inference apparatus according to  claim 2 , wherein
 the neural network model includes an RNN (recurrent neural network), and   the one or more processors sequentially acquire the future latent state through the RNN.   
     
     
         7 . The inference apparatus according to  claim 2 , wherein
 at a focused time, the one or more processors are configured to:
 acquire the control data in time series from the focused time for a predetermined time period; 
 acquire the latent state at the focused time based on the control data at the focused time and an initial value of the latent state; 
 acquire a predicted value of the task at the focused time based on the latent state at the focused time; 
 repeat arithmetic operation to acquire the latent state and acquire the predicted value of the task until the predicted values of the task for the predetermined time period are acquired; and 
 update the time-series control data based on the predicted values of the task for the predetermined time period. 
   
     
     
         8 . The inference apparatus according to  claim 7 , wherein
 at the focused time, the one or more processors acquire the time-series control data at the focused time based on the time-series control data, which is updated before the focused time.   
     
     
         9 . The inference apparatus according to  claim 8 , wherein
 the one or more processors acquire the time-series control data at the next time by shifting the time-series control data updated at the focused time for one period of time.   
     
     
         10 . The inference apparatus according to  claim 8 , wherein
 the one or more processors acquire the time-series control data at the next time based on a maximum value and a minimum value out of the time-series control data updated at the focused time.   
     
     
         11 . The inference apparatus according to  claim 2 , wherein
 the neural network model is a neural network optimized based on at least one of a joint state of the control target acquired based on the latent state and the image of the control target acquired based on the latent state, and data indicating the respective target states in addition to the time series of the task and the control data.   
     
     
         12 . The inference apparatus according to  claim 1 , wherein
 the one or more processors is configured to:   acquire a plurality of the time-series control data;   calculate the losses based on the respective time-series control data; and   update the time-series control data based on the time-series control data, which outputs the minimum loss, out of the calculated plurality of losses.   
     
     
         13 . The inference apparatus according to  claim 1 , wherein
 the one or more processors acquire input data at the next time based on a result executed based on the output time-series control data.   
     
     
         14 . The inference apparatus according to  claim 1 , wherein
 the task indicates a sparse state, and   the one or more processors calculate the loss by comparing the task indicating the sparse state and the time series of the task.   
     
     
         15 . The inference apparatus according to  claim 1 , wherein
 the time series of the task inferred from the future latent state is a time series of a state of a deliverable to be produced by the control target controlled based on the control data.   
     
     
         16 . The inference apparatus according to  claim 1 , wherein
 the calculating calculates the loss, based on the inferred time series of a deliverable to be produced by the task and the data indicating the target state of the deliverable to be produced by the task.   
     
     
         17 . The inference apparatus according to  claim 1 , wherein
 the input data is observed data acquired from observing the control target, and the target state is a target state of a deliverable produced by the control target.   
     
     
         18 . The inference apparatus according to  claim 1 , wherein
 the control target is one of a robot system or a process system.   
     
     
         19 . A training apparatus, comprising:
 one or more memories; and   one or more processors configured to:
 acquire a latent state from input data regarding a control target; 
 acquire a future latent state from a control data and the latent state; 
 infer, from the future latent state, a time series of a task to be executed by the control target to be controlled based on the control data; 
 acquire the latent state from the input data by comparing a state data of the control target based on the latent state and the time series of the task with data indicating respective target states, acquire the future latent state, and update a neural network model inferring the time series of the task. 
   
     
     
         20 . An inference method, comprising:
 acquiring a latent state from input data regarding a control target;   acquiring a future latent state from the latent state and a control data;   inferring, from the future latent state, a time series of a task to be executed by the control target to be controlled based on the control data;   calculating a loss between the time series of the task and data indicating a target state; and   updating the control data based on the loss.